Pipeline
- Execution Lifecycle
Requests are classified as Spike, Bounded, or Architectural, then pass brainstorming, plan approval, and verification gates before a debt sweep.
- Execution Model: Subagent-First, High Fan-Out
The parent agent splits work into small verifiable chunks, dispatches 10 or more subagents when it can, then gathers and synthesizes their reports.
- Confirmations Without Typing
Each human decision gate is asked through the agent tool's own question tool with ready-made options, so an answer is a tap or an Enter.
- Deep Research Workflow
When a choice cannot be settled from the codebase, a cited research report is written between the spike and the implementation plan.
- Executing an ultra-plan/v1 Plan
scripts/ultra-plan-runner.mjs parses plan frontmatter, validates the task DAG, skips idempotent tasks, runs each step, and records failures in a ledger.
- Plan Mirroring: Obsidian and GitHub Issues
Plans are copied one way to Obsidian and GitHub issues, and an issue body is the plan text byte for byte plus a trailer with id, status, and hash.
- Plan Checklist (the to-do list)
The to-do list is a checklist in the plan file, owned by the parent agent, and each checked item cites its command, exit code, and result.
- Pull Request Delivery & Batch Merge
Each session ends in a pull request on its own branch, git writes stay with the parent, and a PR registry merges sessions in dependency order.